blog single

What AI can and cannot do for a service business

Team reviewing an AI assisted workflow on a screen

Published on:

January 16, 2025

TAG:

BY:

Edgeless Team

Read Time:

5 MIN READ

post detail

AI tools are now in every software suite, and most service businesses have tried at least one. The results are uneven. Some teams save hours a week on drafting and summarising. Others add a tool that produces confident text nobody trusts and a review step that costs more than the work it replaced. The difference is usually the use case, not the model.

The useful part is narrow

Language models are good at transforming text that already exists. Summarising a long document, rewriting a paragraph for a different audience, extracting fields from an email, drafting a first version of a routine reply. They are weaker at work that needs a source of truth, a calculation or a judgement about a person.

That distinction matters because the failures are quiet. A summary that misses the one clause that changes the meaning looks as polished as a correct one. A drafted reply that invents a delivery date reads like every other reply. The output is fluent, and fluency is easy to mistake for accuracy.

Where it fits today

Turning meeting notes into a short list of actions with owners.
Drafting a first reply to a routine enquiry from a template and the customer's message.
Extracting names, dates and amounts from documents into a structured record.
Summarising a long support thread before a handover.
Rewriting technical notes for a customer audience.
Classifying incoming messages by topic so they reach the right queue.

Each of these has a human check at the end and a clear input. The tool works on text that a person already produced, and a person reviews the result before it reaches a customer or a system of record. That structure keeps the risk contained.

Where it fails quietly

AI fails when it is asked to know something it was not given. Prices, stock levels, appointment availability, contract terms and the status of a job live in your systems, not in a model's training data. A chatbot that answers those questions without a connection to the source will produce plausible answers that are wrong.

It also fails when the cost of a wrong answer is high and the review is weak. A drafted marketing paragraph is easy to check. A generated quote, a medical instruction or a legal notice needs a qualified person to approve it, and the approval has to be built into the workflow rather than assumed.

Watch for automation bias as well. When a tool produces a draft, reviewers tend to accept most of it. The review becomes a formality, and the errors that survive are the ones nobody expected. Rotate samples to a second reviewer and track corrections to see whether the check is real.

Engineer reviewing AI output against source documents — The reliable pattern is a clear input, a bounded task and a person who signs off before anything leaves the building.

Test a use case before you buy

Pick one task that happens often and takes a predictable amount of time. Collect twenty real examples, including the messy ones. Run them through the tool and score the output against what a competent person would produce. Count the minutes saved and the minutes spent correcting.

If the correction time is close to the original time, the use case is not ready, even if the demo looked impressive. If corrections are small and repeatable, the tool may be worth adopting for that task. The test costs a day and prevents a year of subscription and review overhead.

Include the failure cases in the test. What happens when the input is incomplete, the language is mixed or the document is a scan. Those cases decide whether the workflow needs a fallback to a person, and every production workflow needs one.

Read the data terms

Before sending business data to a tool, find out whether the provider uses it for training, where it is stored and how long it is retained. For customer records, contracts and anything personal, the answer decides whether the tool can be used at all. Put the answer in writing and keep it with the vendor record.

Check who can see the data inside your own team. A shared account on a chat tool means every conversation is visible to everyone with the login. Named accounts with roles cost more and make it possible to remove access when someone leaves.

People still own the output

The team that uses the tool needs to understand what it is doing and where it stops. A short guide with examples of good and bad output is more useful than a policy document. Name the person who approves each kind of output, and make it easy to reject a draft without arguing with the tool.

Keep a record of what was generated and who approved it for anything customer facing. That record protects the business if a question arises later, and it gives the team evidence for improving the prompts and the review steps.

AI is useful when it shortens a task a person already understands. It is risky when it replaces the understanding.
Edgeless Team

A sensible first project

Start with internal drafting. A tool that helps the team write a first version of a proposal section, a job description or a customer update has a low blast radius and an obvious review step. Measure the time saved for a month before expanding.

Then connect one source of truth. A assistant that can read the job status from your system and answer a question about it is far more useful than one that guesses. The connection is the work. The model is a component, and the integration is what makes the answer trustworthy.

Avoid the temptation to launch a customer facing chatbot before the internal workflow is reliable. Customers test edge cases faster than any internal team, and a public failure costs more trust than the time saved.

Wrap the model in a workflow

Most successful AI projects look boring in a demo. A staff member opens a record, presses a button and gets a draft with the source documents attached. They edit it, approve it and the record shows who generated it and who approved it. No conversation, no prompt engineering at the point of use.

That design keeps the tool inside a process the business already understands. It also makes failure visible. When the draft is wrong, the person sees the source and can correct it. When the tool is unavailable, the workflow falls back to the manual step instead of stopping.

What to measure

Measure minutes saved after correction, not minutes saved on the first draft. Measure the share of drafts accepted without edits and the types of edits that recur. A rising acceptance rate on a stable task is evidence that the workflow is working. A flat rate after three months is evidence that the task or the review needs a different design.

Track customer complaints and escalations as well. A tool that saves the team time and increases the number of corrections from customers has moved work, not removed it. Those numbers belong in the same review as the time saved.

We build AI assisted workflows with source connected data, review steps and an audit trail, starting from one task that already has a clear owner.

Explore AI development

The businesses that get value from AI pick a narrow task, measure the correction cost and keep a person accountable for the result, whatever the size of their tool stack. That approach works with today's models and will still work when the next ones arrive.

Operations

Automation

What AI can and cannot do for a service business

Let's Talk About Your Project